From: <pbhagavatula@marvell.com>
To: <jerinj@marvell.com>, Pavan Nikhilesh <pbhagavatula@marvell.com>,
"Shijith Thotton" <sthotton@marvell.com>,
Nithin Dabilpuram <ndabilpuram@marvell.com>,
Kiran Kumar K <kirankumark@marvell.com>,
Sunil Kumar Kori <skori@marvell.com>,
Satha Rao <skoteshwar@marvell.com>
Cc: <dev@dpdk.org>
Subject: [PATCH v2 2/2] net/cnxk: align perfetchs to CN10K cache model
Date: Thu, 24 Feb 2022 21:40:12 +0530 [thread overview]
Message-ID: <20220224161013.4566-2-pbhagavatula@marvell.com> (raw)
In-Reply-To: <20220224161013.4566-1-pbhagavatula@marvell.com>
From: Pavan Nikhilesh <pbhagavatula@marvell.com>
Align perfetchs for CN10K cache model for vWQE in Rx and Tx.
Move mbuf->next NULL assignment to Tx path and enabled it only
when multi segments offload is enabled to reduce L1 pressure.
Add macros to detect corrupted mbuf->next values when
MEMPOOL_DEBUG is set.
Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
---
drivers/event/cnxk/cn10k_worker.h | 13 ++--
drivers/net/cnxk/cn10k_rx.h | 115 ++++++++++++++++++++++++------
drivers/net/cnxk/cn10k_tx.h | 7 ++
3 files changed, 107 insertions(+), 28 deletions(-)
diff --git a/drivers/event/cnxk/cn10k_worker.h b/drivers/event/cnxk/cn10k_worker.h
index d288c66cac..a827a1e422 100644
--- a/drivers/event/cnxk/cn10k_worker.h
+++ b/drivers/event/cnxk/cn10k_worker.h
@@ -118,23 +118,23 @@ cn10k_process_vwqe(uintptr_t vwqe, uint16_t port_id, const uint32_t flags,
uint64_t aura_handle, laddr;
uint16_t nb_mbufs, non_vec;
uint16_t lmt_id, d_off;
+ struct rte_mbuf **wqe;
struct rte_mbuf *mbuf;
uint8_t loff = 0;
uint64_t sa_base;
- uint64_t **wqe;
int i;
mbuf_init |= ((uint64_t)port_id) << 48;
vec = (struct rte_event_vector *)vwqe;
- wqe = vec->u64s;
+ wqe = vec->mbufs;
- rte_prefetch_non_temporal(&vec->ptrs[0]);
+ rte_prefetch0(&vec->ptrs[0]);
#define OBJS_PER_CLINE (RTE_CACHE_LINE_SIZE / sizeof(void *))
for (i = OBJS_PER_CLINE; i < vec->nb_elem; i += OBJS_PER_CLINE)
- rte_prefetch_non_temporal(&vec->ptrs[i]);
+ rte_prefetch0(&vec->ptrs[i]);
nb_mbufs = RTE_ALIGN_FLOOR(vec->nb_elem, NIX_DESCS_PER_LOOP);
- nb_mbufs = cn10k_nix_recv_pkts_vector(&mbuf_init, vec->mbufs, nb_mbufs,
+ nb_mbufs = cn10k_nix_recv_pkts_vector(&mbuf_init, wqe, nb_mbufs,
flags | NIX_RX_VWQE_F, lookup_mem,
tstamp, lbase);
wqe += nb_mbufs;
@@ -182,7 +182,7 @@ cn10k_process_vwqe(uintptr_t vwqe, uint16_t port_id, const uint32_t flags,
cn10k_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf, tstamp,
flags & NIX_RX_OFFLOAD_TSTAMP_F,
(uint64_t *)tstamp_ptr);
- wqe[0] = (uint64_t *)mbuf;
+ wqe[0] = (struct rte_mbuf *)mbuf;
non_vec--;
wqe++;
}
@@ -612,6 +612,7 @@ cn10k_sso_hws_event_tx(struct cn10k_sso_hws *ws, struct rte_event *ev,
ev->sched_type, txq_data, flags);
}
rte_mempool_put(rte_mempool_from_obj(ev->vec), ev->vec);
+ rte_prefetch0(ws);
return (meta & 0xFFFF);
}
diff --git a/drivers/net/cnxk/cn10k_rx.h b/drivers/net/cnxk/cn10k_rx.h
index 65a08e379b..66a35c69f9 100644
--- a/drivers/net/cnxk/cn10k_rx.h
+++ b/drivers/net/cnxk/cn10k_rx.h
@@ -36,6 +36,22 @@
(((f) & NIX_RX_VWQE_F) ? \
(uint64_t *)(((uintptr_t)((uint64_t *)(b))[i]) + (o)) : \
(uint64_t *)(((uintptr_t)(b)) + CQE_SZ(i) + (o)))
+#define CQE_PTR_DIFF(b, i, o, f) \
+ (((f) & NIX_RX_VWQE_F) ? \
+ (uint64_t *)(((uintptr_t)((uint64_t *)(b))[i]) - (o)) : \
+ (uint64_t *)(((uintptr_t)(b)) + CQE_SZ(i) - (o)))
+
+#ifdef RTE_LIBRTE_MEMPOOL_DEBUG
+#define NIX_MBUF_VALIDATE_NEXT(m) \
+ if (m->nb_segs == 1 && mbuf->next) { \
+ rte_panic("mbuf->next[%p] valid when mbuf->nb_segs is %d", \
+ m->next, m->nb_segs); \
+ }
+#else
+#define NIX_MBUF_VALIDATE_NEXT(m) \
+ do { \
+ } while (0)
+#endif
union mbuf_initializer {
struct {
@@ -674,17 +690,73 @@ cn10k_nix_recv_pkts_vector(void *args, struct rte_mbuf **mbufs, uint16_t pkts,
cq0 = (uintptr_t)&mbufs[packets];
}
- /* Prefetch N desc ahead */
- rte_prefetch_non_temporal(CQE_PTR_OFF(cq0, 4, 64, flags));
- rte_prefetch_non_temporal(CQE_PTR_OFF(cq0, 5, 64, flags));
- rte_prefetch_non_temporal(CQE_PTR_OFF(cq0, 6, 64, flags));
- rte_prefetch_non_temporal(CQE_PTR_OFF(cq0, 7, 64, flags));
-
- /* Get NIX_RX_SG_S for size and buffer pointer */
- cq0_w8 = vld1q_u64(CQE_PTR_OFF(cq0, 0, 64, flags));
- cq1_w8 = vld1q_u64(CQE_PTR_OFF(cq0, 1, 64, flags));
- cq2_w8 = vld1q_u64(CQE_PTR_OFF(cq0, 2, 64, flags));
- cq3_w8 = vld1q_u64(CQE_PTR_OFF(cq0, 3, 64, flags));
+ if (flags & NIX_RX_VWQE_F) {
+ if (pkts - packets > 4) {
+ rte_prefetch_non_temporal(
+ CQE_PTR_OFF(cq0, 4, 0, flags));
+ rte_prefetch_non_temporal(
+ CQE_PTR_OFF(cq0, 5, 0, flags));
+ rte_prefetch_non_temporal(
+ CQE_PTR_OFF(cq0, 6, 0, flags));
+ rte_prefetch_non_temporal(
+ CQE_PTR_OFF(cq0, 7, 0, flags));
+
+ if (likely(pkts - packets > 8)) {
+ rte_prefetch1(
+ CQE_PTR_OFF(cq0, 8, 0, flags));
+ rte_prefetch1(
+ CQE_PTR_OFF(cq0, 9, 0, flags));
+ rte_prefetch1(
+ CQE_PTR_OFF(cq0, 10, 0, flags));
+ rte_prefetch1(
+ CQE_PTR_OFF(cq0, 11, 0, flags));
+ if (pkts - packets > 12) {
+ rte_prefetch1(CQE_PTR_OFF(
+ cq0, 12, 0, flags));
+ rte_prefetch1(CQE_PTR_OFF(
+ cq0, 13, 0, flags));
+ rte_prefetch1(CQE_PTR_OFF(
+ cq0, 14, 0, flags));
+ rte_prefetch1(CQE_PTR_OFF(
+ cq0, 15, 0, flags));
+ }
+ }
+
+ rte_prefetch0(CQE_PTR_DIFF(
+ cq0, 4, RTE_PKTMBUF_HEADROOM, flags));
+ rte_prefetch0(CQE_PTR_DIFF(
+ cq0, 5, RTE_PKTMBUF_HEADROOM, flags));
+ rte_prefetch0(CQE_PTR_DIFF(
+ cq0, 6, RTE_PKTMBUF_HEADROOM, flags));
+ rte_prefetch0(CQE_PTR_DIFF(
+ cq0, 7, RTE_PKTMBUF_HEADROOM, flags));
+ if (likely(pkts - packets > 8)) {
+ rte_prefetch0(CQE_PTR_DIFF(
+ cq0, 8, RTE_PKTMBUF_HEADROOM,
+ flags));
+ rte_prefetch0(CQE_PTR_DIFF(
+ cq0, 9, RTE_PKTMBUF_HEADROOM,
+ flags));
+ rte_prefetch0(CQE_PTR_DIFF(
+ cq0, 10, RTE_PKTMBUF_HEADROOM,
+ flags));
+ rte_prefetch0(CQE_PTR_DIFF(
+ cq0, 11, RTE_PKTMBUF_HEADROOM,
+ flags));
+ }
+ }
+ } else {
+ if (pkts - packets > 4) {
+ rte_prefetch_non_temporal(
+ CQE_PTR_OFF(cq0, 4, 64, flags));
+ rte_prefetch_non_temporal(
+ CQE_PTR_OFF(cq0, 5, 64, flags));
+ rte_prefetch_non_temporal(
+ CQE_PTR_OFF(cq0, 6, 64, flags));
+ rte_prefetch_non_temporal(
+ CQE_PTR_OFF(cq0, 7, 64, flags));
+ }
+ }
if (!(flags & NIX_RX_VWQE_F)) {
/* Get NIX_RX_SG_S for size and buffer pointer */
@@ -997,19 +1069,18 @@ cn10k_nix_recv_pkts_vector(void *args, struct rte_mbuf **mbufs, uint16_t pkts,
nix_cqe_xtract_mseg((union nix_rx_parse_u *)
(CQE_PTR_OFF(cq0, 3, 8, flags)),
mbuf3, mbuf_initializer, flags);
- } else {
- /* Update that no more segments */
- mbuf0->next = NULL;
- mbuf1->next = NULL;
- mbuf2->next = NULL;
- mbuf3->next = NULL;
}
- /* Prefetch mbufs */
- roc_prefetch_store_keep(mbuf0);
- roc_prefetch_store_keep(mbuf1);
- roc_prefetch_store_keep(mbuf2);
- roc_prefetch_store_keep(mbuf3);
+ /* Mark mempool obj as "get" as it is alloc'ed by NIX */
+ RTE_MEMPOOL_CHECK_COOKIES(mbuf0->pool, (void **)&mbuf0, 1, 1);
+ RTE_MEMPOOL_CHECK_COOKIES(mbuf1->pool, (void **)&mbuf1, 1, 1);
+ RTE_MEMPOOL_CHECK_COOKIES(mbuf2->pool, (void **)&mbuf2, 1, 1);
+ RTE_MEMPOOL_CHECK_COOKIES(mbuf3->pool, (void **)&mbuf3, 1, 1);
+
+ NIX_MBUF_VALIDATE_NEXT(mbuf0);
+ NIX_MBUF_VALIDATE_NEXT(mbuf1);
+ NIX_MBUF_VALIDATE_NEXT(mbuf2);
+ NIX_MBUF_VALIDATE_NEXT(mbuf3);
packets += NIX_DESCS_PER_LOOP;
diff --git a/drivers/net/cnxk/cn10k_tx.h b/drivers/net/cnxk/cn10k_tx.h
index ec6366168c..695e3ed354 100644
--- a/drivers/net/cnxk/cn10k_tx.h
+++ b/drivers/net/cnxk/cn10k_tx.h
@@ -2569,6 +2569,13 @@ cn10k_nix_xmit_pkts_vector(void *tx_queue, uint64_t *ws,
lnum += 1;
}
+ if (flags & NIX_TX_MULTI_SEG_F) {
+ tx_pkts[0]->next = NULL;
+ tx_pkts[1]->next = NULL;
+ tx_pkts[2]->next = NULL;
+ tx_pkts[3]->next = NULL;
+ }
+
tx_pkts = tx_pkts + NIX_DESCS_PER_LOOP;
}
--
2.17.1
next prev parent reply other threads:[~2022-02-24 16:10 UTC|newest]
Thread overview: 7+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-02-24 13:52 [PATCH 1/2] net/cnxk: optimize Rx pktsize extraction pbhagavatula
2022-02-24 13:52 ` [PATCH 2/2] event/cnxk: align perfetchs to CN10K cache model pbhagavatula
2022-02-24 16:10 ` [PATCH v2 1/2] net/cnxk: optimize Rx pktsize extraction pbhagavatula
2022-02-24 16:10 ` pbhagavatula [this message]
2022-02-24 18:40 ` [dpdk-dev] [PATCH v3 1/2] net/cnxk: optimize Rx packet size extraction jerinj
2022-02-24 18:40 ` [dpdk-dev] [PATCH v3 2/2] net/cnxk: align perfetchs to CN10K cache model jerinj
2022-02-24 20:39 ` [dpdk-dev] [PATCH v3 1/2] net/cnxk: optimize Rx packet size extraction Jerin Jacob
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20220224161013.4566-2-pbhagavatula@marvell.com \
--to=pbhagavatula@marvell.com \
--cc=dev@dpdk.org \
--cc=jerinj@marvell.com \
--cc=kirankumark@marvell.com \
--cc=ndabilpuram@marvell.com \
--cc=skori@marvell.com \
--cc=skoteshwar@marvell.com \
--cc=sthotton@marvell.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).